OR Spectrum (2010) 32:765786

DOI 10.1007/s00291-010-0197-0
REGULAR ARTICLE

Positioning empty containers among multiple ports

with leasing and purchasing considerations
Il-Kyeong Moon Anh-Dung Do Ngoc
Yun-Su Hur

Published online: 10 March 2010

Springer-Verlag 2010

Abstract With the recent development of container transportation, the imbalance

of empty containers among ports has become more serious. We consider the problem
of positioning empty containers. The goal of this study is to propose a plan for transporting empty containers between container ports (terminals) to reduce the imbalance.
There is currently a demand at each port and any backlog of containers is not permitted. The objective is to minimize the total relevant costs such as transportation cost,
handling cost, and holding cost, etc. In this study, we develop a model with respect
to the leasing and purchasing of containers. Mixed integer programming and genetic
algorithms are used to solve the model. A hybrid GA is also proposed to reduce the
computation time while still obtaining an acceptable result.
Keywords Empty container positioning Hybrid genetic algorithm
Mixed integer programming
1 Introduction
The transportation of goods has become more developed with the acceleration of the
world economy and the consequent increase in global trade. One of the most effective
and least expensive commodities for transportation is containers. For inland transportation, there are several options for transporting goods, but for ocean transportation, only
containers and container vessels can be used. A container, which is fully loaded with

I.-K. Moon (B) A.-D. Do Ngoc

Department of Industrial Engineering, Pusan National University, Busan 609-735, Korea
e-mail: ikmoon@pusan.ac.kr
Y.-S. Hur
Urban Creation Research Group, Busan Development Institute, Busan, Korea

123

766

I.-K. Moon et al.

goods from the supplier, is transported across oceans to the destination port and is then
delivered to the customer and unloaded. After unloading, the container is maintained
and stored at the destination port to await another trip. When there is an imbalance
between the number of import containers and the number of export containers, some
ports have a surplus of empty containers while other ports have a deficit. When ports
have a deficit of empty containers, shipping companies must lease or purchase empty
containers, and when ports have a surplus, the empty containers are stored in depots.
Strategically, positioning the empty containers is, therefore, one of the most effective
ways to solve the problem of the imbalance of containers. In 1995, the flow from Asia
to the USA was 4 million 20-foot equivalent units (TEUs) and 3.5 million TEUs in
the reverse direction. By 2005, the annual flow had increased to 12.4 million TEUs
from Asia to the USA and 4.2 million TEUs in the reverse direction. By 2007, the
annual flow was 15.4 million TEUs from Asia to the USA and 4.9 million TEUs in
the reverse direction. (source: http://people.hofstra.edu/geotrans/eng/ch3en/conc3en/
worldcontainerflows.html). Hence, the annual container flow from Asia to the USA
increased by 8.4 million TEUs, in the 10 years from 1995 to 2005 and by 3 million
TEUs in the 2 years from 2005 to 2007. In addition, the imbalance of container flow
between Asia and the USA has also increased. The imbalance in 1995 was 0.5 million
TEUs, that in 2005 was 8.2 million TEUs, and that in 2007 was 10.5 million TEUs.
These numbers show that container flows and imbalance are continuously increasing. The situation is similar for container flows between Asia and Europe. Several
approaches have been used by shipping companies to manage the imbalance of container flows and satisfy the demand. Among them, empty container positioning and
leasing are widely considered. There have been many studies on empty container
positioning, some of which focus on the leasing of containers, but they concentrate on
long-term leasing only.
In this paper, we consider the problem of ocean positioning planning. In addition,
to meet the demand, we also consider short-term leasing and purchasing factors. In
long-term leasing, we can consider the leased container to be an owned one, assuming
that the lease is for a considerably long period of time. However, in short term leasing,
the lease duration is short (in several periods), and container handling comes into the
picture when the container is being returned. The objective of this study is to minimize the total relevant costs, including transportation cost, handling cost, inventory
holding cost, leasing cost, and purchasing cost. The remaining content of the paper
is organized as follows: in Sect. 2, we develop the mathematical model. Two solution
algorithms are presented in Sect. 3. Section 4 illustrates the computational results, and
conclusion is presented in the final section.

2 The model
2.1 Literature review
Recently, many studies have been carried out concerning empty container positioning.
Crainic et al. (1989) considered the multi-commodity location problem with balancing requirements. They proposed models for multi-commodity capacitated location

123

Positioning empty containers among multiple ports

767

problems (MCLB) with an inter-depot balancing requirement. The decision variables

of MCLB included a set of binary variables which decides on the opening or closing
of depots and a set of continuous variables that presents the empty container flows
among supply customers, depots, and demand customers. The objective function is
to minimize the total cost, which involves the cost of opening the depot and transportation costs. The demand for empty containers must be satisfied. There have also
been many studies carried out that attempt to solve the MCLB problem. Crainic and
Delorme (1993) developed dual-ascent procedures for the proposed model. Crainic
et al. (1993) solved the problem by using the Tabu search procedure. Gendron and
Crainic (1997) presented the parallel branch and bound approach, which is based on a
dual-ascent procedure previously proposed by Crainic and Delorme (1993). Gendron
et al. (2003) also solved the problem using the Tabu search procedure, but they used
the slope scaling method to find the starting solution. Li et al. (2004) studied the management of empty containers in a port with stochastic demand. Their analysis is based
on the multistage inventory problem and the Markov decision processes with discrete
time. They focused on the optimization of the pair-critical policy (U , D). In this policy, if the number of empty containers at a port is less than U, the empty containers
are imported up to the amount of U; and if it is more than D, the empty containers
are exported down to the amount of D. Recently, Li et al. (2007) have since advanced
the problem for the multi-port. Shen and Khoong (1995) proposed a decision support
system (DSS) for empty container distribution planning. The DSS is based on the network optimization models. In the network, they considered the leasing-in, off-leasing,
positioning-in, and positioning-out at a port. The problem is decomposed into three
levels, namely the terminal (port) planning, intra-regional planning, and inter-regional
planning. They considered the single type of containers.

2.2 Problem definition

First, the activity of a shipping company is described by considering the following scenario. At a port, the shipping company receives orders from customers.
After receiving an order, the company transports empty containers to the customer.
The customers cargo will then be loaded into the containers. These full containers are then transported to the terminal and carried by a vessel to the destination port. At the destination port, the full containers are delivered, the cargo is
unloaded, and the container is transported back to the depot. After being checked and
maintained, these empty containers are ready for a new trip. When the full containers are transported to other ports, they are exported whereas if the full containers are received from another port, they are imported. If the number of imported
containers is either less than or greater than that of exported containers, there will
be an imbalance. It can be, therefore, seen that if the number of exported containers is greater than that of imported containers, there will be a shortage of
empty containers. For the reverse situation, there would be a surplus of empty
containers. At the surplus ports, the surplus containers will be stored in the depot
and the cost for holding these containers must be paid. At the ports, where there
is a shortage of containers, empty containers must be leased or purchased to

123

768

I.-K. Moon et al.

t
Port I

t+1
Returned
Containers

Port J

Port I

t+2
Leasing
Containers

Port J

Port J

Purchased
Containers

Scrapped
Containers
Inventory of Owned
Containers

Port I

Inventory of Leased
Containers

Demand

Owned Inventory

Leased Inventory

Full Container

Devanning Container

Owned-Container Positioning

Leased-Container Positioning

Fig. 1 The flow of containers between ports in multi periods

fulfill demand. The shipping company needs to find a way to reduce these costs.
One solution involves empty containers being transported from ports with surplus
containers to ports with a shortage. Transporting empty containers from port to port
is termed empty container positioning.
When empty containers are positioned, the holding cost, leasing cost, and purchasing cost can be reduced, but there is a transportation cost. Moreover, because of the
fluctuation in demand, a port may have a surplus at one time and a shortage another.
Therefore, the shipping company must decide how many empty containers should be
positioned, leased, and purchased at each port at various times. We therefore aim to
propose a mathematical model to solve this problem.
Figure 1 graphically describes the situation. Continuous arrows represent the flow
of empty containers, while dashed arrows represent the flow of full containers. At
any port, the demand is fulfilled by empty containers in the inventory. There are two
kinds of containers: owned containers and leased containers. Therefore, the inventory is divided into: the owned container inventory and leased container inventory.
A number of continuous arrows connect the two types. In Fig. 1, node I denotes
an owned-container inventory, while node L denotes a leased-container inventory.
The arrows connecting the inventory of a port to itself (I I and L L at the
same port) in the next period represent the flow of inventory to the next period.
The arrows connecting the inventory of one port to that of another port (I I
and L L for different ports) represent the flow of positioned empty containers.

123

Positioning empty containers among multiple ports

769

The dashed arrows from the inventory of one port to the demand node of another
port in the next period (I D and L D) represent the flow of full containers, which are used to satisfy the demand. However, in the mathematical model,
this does not mean that the full containers will arrive at the destination port after
leaving the departure port in one period. They may arrive after several periods,
depending on the transportation time. These flows represent the transport activity,
and there are several kinds of costs for this activity. We divide the costs into fixed
and variable costs. Fixed costs do not depend on the number of transported containers. The operation cost of a vessel is an example. This cost depends on the
locations of the original and destination ports. The administration cost is also in
this category. Variable costs depend on the number of transported containers. There
is also a cost for loading and unloading containers at a certain port. This kind
of cost is referred to as a handling cost and has both fixed and variable components.
As mentioned above, after the full containers have been delivered to customers,
they are unloaded. The containers are then checked, maintained, and returned to the
depot for the next trip. In this paper, these containers are referred to as devanning
containers (the word devanning is used by Hyundai Merchant Marine Company).
In Fig. 1, the arrows connecting the demand nodes to the inventory nodes (D I and
D L) at the same port in the next period represent the flow of devanning containers.
We assume that the devanning time is one period. This means that full containers are
emptied and returned to the inventory in the next period after they have arrived at the
destination port.
There are a number of arrows that point away from or toward owned-container
inventory nodes. The arrows pointing to the owned-container inventory in a certain period represent the flow of purchased containers. These purchased containers
are bought from suppliers and they are first used in that period. The cost relating to this flow is the purchasing cost, and it depends on the container type. The
arrows that point away from the owned-container inventory in a period represent
the flow of scrapped containers. Scrapped containers are those that can no longer
be used. These containers may have expired or be used for another purpose. There
are also a number of arrows that point away from or toward leased-container inventory nodes. The arrows that point away represent the flow of returned containers.
In leasing a container, the shipping company has to pay for activities that do not
depend on the number of leased containers, such as administration. Hence, the leasing cost should include both fixed and variable costs. After the containers have been
leased from the leasing company and used for one or more trips, they are returned
to their owners. These containers are termed returned containers. The arrows that
point toward leased-container inventory nodes represent containers leased in that
period.
In the real world, there are several kinds of container types such as 20, 40 ft, dry,
open-top or refrigerated containers. Thus, in this study, we consider multiple types
of containers. Moreover, the short-term leasing of containers is considered, which
means that containers can be leased and then returned to another port determined by
the lessor. This short-term leasing consideration has not been considered in previous
studies.

123

770

I.-K. Moon et al.

2.3 The mathematical model

A number of assumptions have been made as follows:
The demand must be satisfied, and no backlog is permitted.
The total capacity of owned vessels and rented vessels is enough to transport the
full containers for a certain period.
The route of the vessels is not considered.
The full containers will be emptied after they have arrived at the destination port
within one period.
There is no limitation to the number of purchased containers within a period.
We propose a mathematical model as follows:
Indices:
t: time period, t = 1, 2, 3, . . . , T
i, j, k: port, i, j, k = 1, 2, . . . , P
v: type of containers, v = 1, 2, . . . ,V
Parameters:
Inventory
K iv : maximum inventory level of type v containers at port i
Hiv : unit cost of holding type v containers at port i.
Leasing
LKiv :
LFiv :
LViv :
k :
RKiv

capacity of leasing type v containers at port i

fixed leasing cost of type v containers at port i
variable leasing cost of type v containers at port i
maximum number of type v containers, which are leased at port k and that
will be returned to port i

Transportation
TFi1j : fixed cost for transporting containers from port i to port j by owned vessels
TVi1jv : variable transportation cost of type v container from port i to port j by owned
vessels
TFi2j : fixed cost for transporting containers from port i to port j by rented vessels
TVi2jv : variable transportation cost of type v container from port i to port j by rented
vessels
TKi1jt : capacity of transportation from port i to port j by owned vessels in period t
TKi2jt : capacity of transportation from port i to port j by rented vessels in period t
TIi j : transportation time from port i to port j.
Handling
RF1ji : fixed handling cost of containers at port i, which are transported from port j
by owned vessels

123

Positioning empty containers among multiple ports

771

RV1jiv : variable handling cost of type v containers at port i, which are transported
from port j by owned vessels
RF2ji : fixed handling cost of containers at port i, which are transported from port j
by rented vessels
RV2jiv : variable handling cost of type v containers at port i, which are transported
from port j by rented vessels
Purchasing
PCiv : unit cost of purchasing type v containers at port i.
Others
Di jvt : demand for type v containers at period t, which should be transported from
port i to port j
SCivt : number of scrapped containers at port i at period t, which are type v
TEUv : changing rate from container type v to TEU.
Decision variables:
Inventory
Iivt : inventory level of type v owned containers at port i at the end of period t
k : inventory level of type v containers at port i in period t, which are leased at
LIivt
port k
Transportation
xiejvt : number of type v empty containers (owned containers) leaving in period t for
being transported from port i to port j
f
xi jvt : number of type v full containers (owned containers leaving in period t for
being transported from port i to port j
:
xielk
jvt number of type v empty containers (leased at port k) leaving in period t for
being transported from port i to port j
f lk
xi jvt : number of type v full containers (leased at port k) leaving in period t for being
transported from port i to port j
f i1jvt : number of type v containers, leaving in period t for being transported from
port i to port j by owned vessels
f i2jvt : number of type v containers, leaving in period t for being transported from
port i to port j by rented vessels

1, if there is one or more vessels leaving in period t for transporting

si1jt : containers from port i to port j

0, otherwise

1, if there is one or more rented vessels leaving in period t for

2
si jt : transporting containers from port i to port j

0, otherwise

123

772

I.-K. Moon et al.

Leasing
yivt : number of type v containers, which are leased at port i in period t
k : number of type v containers, which are leased at port k, and will be returned to
wivt
port i in period t

1, if we lease type v containers at port i in period t
u ivt :
0, otherwise
Purchasing
pivt : number of type v containers purchased at port i in period t.
The mathematical model:
Min Z =

P
P



T


RF1ji s 1ji,tTI ji +

i=1 j=1, j=i t>TI ji

RF2ji s 2ji,tTI ji
P
P
T




+ TFi2j si2jt +

I 
T
V 


I 
T
V 


TFi1j si1jt +

TVi1jv f i1jvt

TVi2jv f i2jvt

Hiv Iivt +


k
LIivt

(LFiv u ivt ) +

i=1 v=1 t=1

1
RV1jiv f jiv,tTI
ji

i=1 v=1 t=1

2
RV2jiv f jiv,tTI
ji

i=1 j=1, j=i t=1

P 
T
P 
V 


k
LVkv LIivt

k=1 i=1 v=1 t=1

P 
T 
P
P
V 






flk
LVkv xielk
jvt + x i jvt TIi j
k=1 i=1 j=1, j=i v=1 t=1

T
I 
V 


PCiv pivt

i=1 v=1 t=1

Subject to
Iivt = Iiv,t1 +

P
T


j=1, t>T I ji
j=i

P

j=1,
j=i

123

x ejiv,tT I ji +

P


T


j=1, t>TI ji +1
j=i

x jiv,tTI ji 1


f
xiejvt + xi jvt + pivt SCivt i, v, t

(1)

Positioning empty containers among multiple ports

773

k
k
LIivt
= LIiv,t1
+

P


P


T


x elk
jiv,tTI ji +

j=1, t>TI ji
j=i

P


T


j=1, t>TI ji +1
j=i

f lk
x jiv,tTI ji 1


f lk
k
xielk
jvt + x i jvt wivt i, v, t, k  = i

(2a)

j=1,
j=i

LIiivt = LIiiv,t1 +

T
P 


x eli
jiv,tTI ji +

j=1, t>TI ji
j=i

P


P


T


j=1, t>TI ji +1
j=i

f li

x jiv,tTI ji 1


f li
i
xieli
+
x
jvt
i jvt + yivt wivt i, v, t

(2b)

j=1,
j=i

xi jvt +

P


f lk

xi jvt = Di jvt i, j = i, v, t

(3)

k=1

Iivt +

P


k
LIivt
K iv i, v, t

(4)

k=1

f i1jvt + f i2jvt

yivt LKiv u ivt i, v, t

k
k
wivt
RKiv
i, v, t, k4
P


f
f lk
xielk
i, j = i, v, t
= xiejvt + xi jvt +
+
x
jvt
i jvt

(5)
(6)
(7)

k=1
V


V



f i1jvt TEUv TKi1jt Si1jt and
f i2jvt TEUv

v=1

v=1

TKi2jt

si2jt

i, j = i, t

f
f lk
k
k
1
2
xiejvt , xi jvt , xielk
jvt , x i jvt , yivt , wivt , Iivt , LIivt , f i jvt , f i jvt ,

u ivt = {0, 1}

(8)
pivt 0, and si1jt , si2jt ,
(9)

The objective function is to minimize the sum of the handling cost, transportation
cost, holding cost, leasing cost, and purchasing cost. Constraints (1), (2a), and (2b) are
inventory balance constraints. The following applies to the owned-container inventory:
The number of empty containers entering the inventory is the sum of the number
of empty containers in the previous period (period t 1), the number of empty
containers received in the current period, the number of full containers received in
the previous period, and the number of containers purchased in the current period

123

774

I.-K. Moon et al.

The number of empty containers leaving the inventory is the sum of the number of
empty and full containers transported to other ports and the number of scrapped
containers
The situation is similar for the leased inventory. The number of empty leased containers entering the inventory is the sum of the number of empty leased containers in
the previous period (period t 1), the number of empty leased containers received in
the current period, the number of full leased containers received in the previous period,
and the number of newly leased containers. The number of empty leased containers
leaving the inventory is the sum of the number of empty and full leased containers
transported to other ports and the number of returned containers that were leased in
previous periods.
Constraint (3) ensures that demands are always satisfied. Constraint (4) avoids the
inventory level that exceeds the maximum inventory level. Constraints (5) and (6)
limit the number of newly leased containers and the number of returned containers,
respectively. Constraints (7) and (8) relate to the vessel. Constraint (7) balances the
total number of transported containers. In the expression of this constraint, the lefthand side represents the total number of transported containers, which is expressed
in terms of the vessel, while the right-hand side represents the total number of owned
and leased containers. Constraint (8) ensures that the number of transported containers
does not exceed the vessel capacity. The vessel capacity is defined in terms of TEU
while the weight capacity is not considered. As in many papers on empty container
movement, especially those written by Crainic et al. (1989, 1993) and Crainic and
Delorme (1993), the variables that relate to the flow of empty containers are considered as continuous variables in this paper. Moreover, the values of these variables are
large and can be rounded to the nearest integer.

3 LP-based genetic algorithm and hybrid genetic algorithm

The above mathematical model is a Mixed Integer Programming (MIP) model. However, the described problem is conjectured as an NP-hard problem and it has been
shown that it is time-consuming to solve this type of a flow model. We therefore
develop two meta-heuristic algorithms.
Drawing from the work of Nemhauser and Wolsey (1988), we can present the MIP
problem in a general form as follows:
(MIP) Min cx + hy
Subject to

Ax + By D
x {0, 1}, y 0

If the value of x = x can be fixed, the problem becomes

(LP(x))

Min
Subject to

123

hy + C
By D A x
y0

where C = c x

Positioning empty containers among multiple ports

775

The LP(x)
is a linear programming problem, which can be easily solved by using a
range of optimization software. If LP (x)
is feasible and its solution is y , then (x,
y )
will be a feasible solution of the MIP problem. If Z MIP is the optimal value of the MIP
is the optimal value of the LP(x)
problem, then Z LP (x)
Z MIP .
problem, and Z LP (x)
When a meta-heuristic is used to generate the value of integer variables, then solving
the LP problem gives the value of continuous variables. The optimum solution of the
MIP problem is denoted as (x , y ). If a value xk exists so that xk = x in the generated
set, then the algorithm can provide us with an optimum solution for the MIP problem.
However, the optimum solution cannot always be found using a meta-heuristic method;
in most cases, this method can only provide the best solution among the available feasible solutions. In this paper, a genetic algorithm, which is a well-known meta-heuristic
method, is used to determine the values of integer variables such that the corresponding
solution comprising a set of continuous and integer variables is near optimal.
There have been several papers on the combination of a GA with Linear Programming (LP). Lou et al. (2001) proposed an integrated method to solve the general MIP
problem and applied it to production planning and scheduling in a batch (PPS) problem. They integrated ordinal optimization and LP together with a GA. In this approach,
the chromosome represents the solution of integers and continuous variables. When a
chromosome is created, its feasibility is checked. If the solution according to a certain
chromosome is infeasible, that chromosome must be re-created. This is a disadvantage
since it may take time to re-create chromosomes. Earlier, Rajasekharan et al. (1998)
also considered combining a GA with MIP to solve the facility layout problem (FLP) in
a flexible manufacturing system (FMS). They suggested using a GA to find the values
of some integer variables, and then to find the values of the remaining integer variables
and continuous variables by applying an LP-based branch and bound procedure. The
disadvantage of this method is that in the first phase, a chromosome cannot give the
values for all integer variables. Therefore, it may take a long time in the second phase
to execute the LP-based branch and bound procedure, instead of implementing the LP.
Pedroso (1998) proposed a method for solving this MIP problem, in which the values of integer variables are determined by an evolutionary algorithm and the values of
continuous variables are then determined by solving the LP problem. He also pointed
out a way to address the infeasibility of the LP problem. In this paper, we propose two
approaches to combine a GA with an LP. In the first approach, we apply the algorithm
proposed by Pedroso (1998), while making a small change to the manner in which the
infeasibility of the LP problem is handled. In Pedrosos algorithm, the chromosome
that makes the relevant LP problem infeasible (which we refer to as an infeasible
chromosome, as opposed to a feasible chromosome) is accepted and measured by
considering its proximity to the feasible region. We, however, have chosen to apply a
procedure called the LP-based GA, which adjusts the infeasible chromosome so that
it becomes feasible. Our second approach-hybrid GA uses a heuristic approach to find
a feasible solution, rather than the optimal solution of the LP problem. The algorithm
that we propose can therefore be described as follows:
1. Generate a set of values of x: (x1 , x2 , . . ., xn ) by a meta-heuristic method.
2. Check the feasibility of the LP(xi ) problem. If the LP(xi ) problem is feasible,
it is solved (heuristic procedure is executed in the second approach) to obtain

123

776

I.-K. Moon et al.

yi and (xi , yi ) is a feasible solution of the MIP problem. Otherwise, we modify

xi so that it makes the LP problem feasible.
3. We then select the feasible solution that provides the smallest objective function
value of the MIP problem. We consider this solution to be the best feasible solution
found by the algorithm.

3.1 LP-based genetic algorithm

Here, we do not explain the theory of GAs. It is more relevant to this study to discuss
how GAs are used to solve our model. The following section presents the structure
of chromosome, fitness function, and the GAs operators and how the experiment is
conducted.
a. Encoding
In the mathematical model, there are three sets of binary variables, namely,
si1jt , si2jt , and u ivt . Decoding each chromosome gives us the values of these variables.
The chromosome is coded as follows:
The chromosome is a string that has three substrings.
Each substring represents the value of the corresponding variables in a set. The
first substring represents the value of si1jt , the second substring represents the value
of si2jt and the third substring is used for u ivt .
Each component (or each gene) of a substring is symbolized by a variable. We can
illustrate how a set of variables is mapped to a substring by using the substring that
is represented by the set of variables si1jt as an example. From left to right along the
1 , the second gene is represented by
substring, the first gene is represented by s111
1 , the T th gene is represented by s1 , the (T + 1)th gene is represented by s 1 ,
s112
121
11T
and the last gene is represented by s 1P P T . Therefore, the length of each substring
determines the number of variables in its corresponding set. In detail, the formula
to compute the chromosome length with respect to P, V, and T is 2 P 2 T + P V T .
The value of each gene is 0 or 1.

The following example is the string for the case where the number of ports is 2, the
number of container types is 2, and the number of period is 4.
1 = 0, s 1 = 1,
From the chromosome presented in Fig. 2, we can see that s111
112
1 = 1, s 2 = 0, s 2 = 1, u
=
1
and
u
=
0.
In
order
to
create
the initial pops223
111
223
112
224
ulation, we assigned the random value (0 or 1) to each gene.
1
0

2
1

15
1

s1ijt

Fig. 2 Chromosome coding

123

16
1

1
0

2
0

s2ijt

15
0

16
1

1
1

2
1

uivt

15
0

16
0

Positioning empty containers among multiple ports

777

b. Fitness function
In the maximization problem, the objective function value can be directly used for
the fitness. However, our mathematical model is a minimization model. Therefore,
if the objective function value of a chromosome is higher, that chromosome shows
less fitness. In this research, the following formula is used to calculate the fitness.
Fitness(i) =  N

MaxZ Z (i)

k=1 [MaxZ

Z (k)]

MaxZ = Max = 1, . . . , N (Z (k)), N : population size

k

where Z (i) is the objective function value of chromosome i.

The denominator of the above formula will be zero when all chromosomes are
identical. In this case, the GA is converged and the solution will be stopped. However,
this case is very unlikely.
c. The solution algorithm
In this section, we discuss how a GA is applied to solve the problem.
Step 1. Generate the initial population.
In this step, a number of chromosomes are created for the first generation. As
mentioned above, a value is randomly assigned to each gene of each chromosome. Subsequently, the feasibility of the solution must be checked. If
the solution given by the chromosome is infeasible, the values of some of the
genes must be changed. The infeasibility appears when demand is higher than
the total capacity of the owned and rented vessels. We have two situations as
follows:
Case 1: si1jt = 1, si2jt = 0(i = j), and demand is higher than the capacity
of the owned vessel; we set si2jt =1.
Case 2: si1jt = 0, si2jt = 0(i = j), and there is a demand from port i to
port j in period t; we set si1jt =1 and continue to check as the first case.
Step 2. Calculate the objective function value and fitness.
In this step, each chromosome represents the value of the binary variables.
Therefore, we will solve the LP(x)
for each chromosome. The feasibility of
the LP(x)
is guaranteed because we assume that there is enough vessel capacity for the demand and there is no limit in the number of purchased containers.
In this research, LINGO is used to solve this LP problem. After solving the
problem, a solution to the continuous variables can be found and the objective
function value and the fitness of each chromosome can easily be calculated.
Step 3. Update the solution and check for the stopping criteria.
In the population, the current best chromosome is chosen and its objective
function value is compared with that of the best chromosome that has been
found so far. For easy understanding, the term the current best chromosome
is used to refer to the chromosome that gives the smallest objective function
among a certain population, and the term the best chromosome is used to
denote the chromosome that gives the best solution found since the beginning
of the GA execution. If the best chromosome has an objective function value

123

778

I.-K. Moon et al.

greater than that of the current best chromosome, the best chromosome is
updated.
Subsequently, the stopping criterion should be checked. The criterion might
either demand that a maximum number of generations has been reached or
that there is no improvement after a number of consecutive generations. In
this research, we use the first criterion.
If the stopping criterion is satisfied, the best chromosome provides the solution. Otherwise, we go to step 4.
Step 4. Create a new generation.
In this step, the GA operators are performed to create a new generation. The
procedure is outlined in the following steps:
Cloning
Selection and crossover
In this research, the roulette wheel and tournament are used for the experiment on GA parameters. After the experiment, the tournament is selected.
For crossover, since the string of each chromosome has three substrings,
a random crossover operator will be performed on each of them. We
based our decision to use random crossover operators on the results of
experiments on GA parameters. Each of the childs genes is randomly
inherited from one of two parent chromosomes.
Mutation
In this research, a uniform method is used in which the value of a gene
is changed from 0 to 1 or from 1 to 0 if it is selected for mutation.
Checking the feasibility of the new chromosome
The population size of the new generation is the same as that of the old
generation.
Return to step 2.
3.2 Hybrid genetic algorithm
As mentioned above, the LP-based GA is able to solve larger problems than the MIP
can. However, it must solve an LP problem for each chromosome. While solving a
single LP problem is not time consuming, solving an extended series of LP problems
is. The LP problem needs to be solved for each chromosome because the chromosome
fitness needs to be more precisely evaluated. Unfortunately, it is seen that this process
takes too much time in the computational experiment section. Therefore, we propose
using a heuristic method rather than an optimization software to solve the LP problem.
Because the heuristic method can find a feasible solution of the LP problem instead of
the optimal solution, it reaches a conclusion more rapidly though less precisely than
the LP-based method. The procedure for the heuristic method is as follows:
a. Heuristic algorithm for finding a feasible solution for the LP problem
Step 1. Assign the empty containers to fulfill the demand.
In this step, we try to fulfill the demand based on the number of on-hand
empty containers. The procedure of this step is as follows:

123

Positioning empty containers among multiple ports

779

Begin

Calculate the number of

available containers.

Available
containers meet the
total demand?

N
The demand will be fulfilled
by:
- Leased containers,
- Owned containers,
- Newly leased and purchased
containers

Owned containers
meet demand?

- Owned containers and leased

containers will be used in
satisfying demand.
- As many as possible of the
remaining leased containers will
be returned.

- Return as many as
possible of the leased
containers.
- The remains will be used
to fulfill the demand.

End

Fig. 3 Flowchart of step 1

Step 1.1. Calculate the number of available containers.

Step 1.2. If owned empty containers can satisfy the total demand,
Return as many as possible of the leased containers.
Assign remaining containers to fulfill the demand. The leased containers have a higher priority for assigning. An attempt is made to
assign the leased containers to the destination to which they can be
returned.
Go to step 1.5
Otherwise, go to the next step.
Step 1.3. If available containers can satisfy the total demand, then
Owned containers and leased containers are used to fulfill demand.
As for the previous step, an attempt is now made to assign the leased
containers to the destination which they can be returned.
As many as possible of the remaining leased containers will be
returned.
Go to step 1.5.
Otherwise, go to the next step.
Step 1.4. All the available containers will be assigned to fulfill part of
the demand. The remaining demand will be satisfied by newly leased and
purchased containers. As for the above step, an attempt is made to assign
the leased containers to the destination to which they can be returned.
Step 1.5. End of step 1.
The flowchart for this step is shown in Fig. 3

123

780

I.-K. Moon et al.

Step 2. In this step, all ports are checked to ensure that they have all been considered.
If all ports have been considered, we go to the next step. Otherwise, the next
port is considered and we return to step 1.
Step 3. Position empty containers.
After fulfilling the demand at all ports, it can be easily determined which ports
have a surplus of empty containers. With these ports, the empty containers
will be positioned to other ports with random amounts.
The procedure is as follows:
Step 3.1. Calculate the total demand and estimate the number of available
containers in the next period. The available containers in the next period
include the surplus empty containers in the current period, the empty containers that will arrive in the next period, and the full containers that will
arrive in this period.
Step 3.2. If the number of available containers is not greater than the total
demand, we proceed directly to step 3.3. Otherwise, we first calculate the
number of containers that are available for positioning by subtracting the
total demand from the number of containers available in the next period.
The remaining containers will be assigned for positioning to other ports
with random amounts.
Step 3.3. If all ports have been considered, we stop and go to step 4.
Otherwise, the next port is chosen and we return to step 3.1.
The flowchart of step 3 is shown in Fig. 4
Step 4. If all periods have been considered, we stop. Otherwise, we return to step 1.
b. The solution algorithm
The above heuristic algorithm can be applied to approximate the optimal solution of
the LP problem for each chromosome and the GA can be executed to solve the IP
problem. It is noted that the solution obtained by the heuristic algorithm is a feasible
one, rather than an optimal one. We therefore need to moderately alter the solution
algorithm. In the LP-based GA approach, there is only one current best chromosome.
However, we do not estimate exactly the fitness of the chromosome in the hybrid-GA.
Hence, instead of using the best chromosome, a set of good chromosomes is considered and the best chromosome among them is chosen after we solve the LP problem for
each of them using optimization software (Lou et al. 2001). The set of good chromosomes includes some chromosomes that have a smaller objective function value than
other chromosomes generated since the beginning of the GA execution. Therefore,
in each generation, after calculating the fitness of each chromosome, the set of good
chromosomes is updated. Note that the size of this set does not change. Thus, if one
or more chromosomes in the current population are better than the chromosomes in
the set of good solutions, they are added to the set and the worst chromosomes in the
set are omitted.
After the GA execution is complete, the set of good chromosomes is known, and
LINGO software is used to solve the LP problem for each chromosome. The chromosome that gives the best objective function value is selected and the best solution is
derived from this chromosome.

123

Positioning empty containers among multiple ports

781

Begin

Calculate the total demand and

estimate the number of available
containers in the next period.

Available
containers meet
the total demand?

Calculate the number of containers that

are available for positioning by
subtracting the total demand from the
available containers in the next period.

N
Randomly assign the number of
containers that are available for
positioning to transport to other ports.

Consider all
ports?

Y
End

Fig. 4 Flowchart of step 3

The solution algorithm of the hybrid GA is similar to that of the LP-based GA and
is presented as a flowchart in Fig. 5. First, chromosomes are generated for the initial
population. The fitness of each chromosome is then calculated. Objective function
values that are derived from chromosomes are used to compute the fitness of each
chromosome. In this step, rather than using LINGO to solve the LP problem, we use
a heuristic algorithm, which has been mentioned above. The next step is updating the
set of good chromosomes. In this step, rather than updating the best chromosome,
we update the set of good chromosomes. The objective function value of each chromosome (in the population of the current generation) is compared with the values of
chromosomes in the set of good chromosomes. Better chromosomes are added to the
set of good chromosomes, while the worst ones are removed. Afterward, the stopping criterion is checked. If the stopping criterion is not satisfied, a new generation
is created by GA operators such as crossover and mutation. Otherwise, GA execution
is stopped and the next step is to find the best chromosome among those in the set
of good chromosomes. The chromosome that gives the best objective function value
is selected. In this step, LINGO is used to solve the LP problem in order to find the
objective function value.
4 Computational results
Initial experiments are conducted to show the performance of the two proposed algorithms. Three problems were generated with a range of sizes, as presented in Table 1.

123

782

I.-K. Moon et al.

Begin

Generate a population.
Solve LP part for each
chromosome by heuristic.

Update the set of good chromosomes.

Is the stopping
criterion
satisfied?

Solve the LP part for each chromosome in

the set of good chromosomes.
Obtain the best
solution.

N
Create a new generation by crossover
and mutation.

End

Fig. 5 Flowchart of the hybrid GA

Table 1 Problem size for initial experiments
Problem
instance

Problem size

Number of variables
Number
of periods

Integer
variables

Continuous
variables

Number of
Constraints

Number
of ports

Number of
container types

Total
variables

I-1

240

3,744

3,984

1,344

I-2

24

720

11,232

11,952

4,032

I-3

30

2,100

49,800

51,900

13,800

All the experiments were conducted on a computer having an Intel Pentium 4 3.4 GHz
processor and 1 GB RAM. LINGO was used to solve the LP part. Table 2 shows the
comparison between all three algorithms. The computation times are given in seconds.
A comparison of the objective values is provided in the last column of the table, and the
differences in objective values are expressed as percentages. These initial experiments
on the computation times and the problem size have shown that the results obtained
by the LP-based GA and by MIP are approximately the same for problems of small
size, and that the LP-based GA can solve the problem with the largest size, while MIP
cannot. However, the computation time is too long. Note that this study targets the
operations of merchant marine companies, and taking decisions on the repositioning
of empty containers is a part of their day-to-day operations. Therefore, we restrict the
acceptable time to be less than a day. If the computation time is longer than a day, we
can safely say that it is too long. LINGO is used for each chromosome to solve the
LP problem. Therefore, in each run, LINGO is used 40 200, i.e., 8,000 times (40 is
the population size and 200 is the maximum generations). Even though the time for
solving each problem is short, the number of times LINGO is used is excessive. For

123

Positioning empty containers among multiple ports

783

Table 2 Comparison between MIP, the LP-based GA, and the hybrid GA
Problem
instance

MIP

LP-based GA (s)

Hybrid GA (s)

Solution
(% penalty)

I-1

<1 min

1,801

84

0.81% (vs. MIP)

I-2

3,105 s

9,807

228

0.75% (vs. MIP)

I-3

>3 days, cannot obtain result

52,051

1,056

0.81% (vs. LP-based)

0.75% (vs. LP-based)
0.56% (vs. LP-based)

this reason, use of the LP-based GA is too time consuming. In Table 2, a considerable
improvement in the computation time is seen for the hybrid GA. The computation
time is reduced by more than 20 times for the smallest problem and about 50 times for
the largest. The objective function values of the hybrid GA are greater than those of
the LP-based GA, but the differences are no more than 1%. This is an acceptable error.
The hybrid GA, therefore, demonstrates efficiency in solving the larger problems.
We also randomly generated 20 problems that had different relevant costs, maximum inventory level, leasing capacity, demand, and other parameters. The number of
ports and the number of container types are randomly generated in the ranges of [3,
10] and [4, 8], respectively. The numbers of periods in weeks are 13, 26, 39, and 52
(corresponding to the number of quarters in a year). The problem sizes are shown in
Table 3. For each problem, the LP-based GA and the hybrid GA are run five times.
The comparison among MIP, the LP-based GA, and the hybrid GA are presented in
Table 4.
In Table 4, the cells containing a symbol indicate that the corresponding problems cannot be solved within a day and results cannot be obtained. The computation
times for the LP-based GA and the hybrid GA are the average computation times for
five runs. The percentage penalty for the hybrid GA is calculated using the average
objective value of five runs. It is seen that the hybrid GA is the superior method in terms
of computation time for large problems. Moreover, when we compare the objective
values, the percentage difference between the hybrid GA result and the results of the
other methods (MIP and the LP-based GA) does not exceed 2 %. Our results present
five distinct problem instances that can be compared. Hence, to make the evaluation
more convincing, we extend it by looking at the lower bound of each problem instance.
In this paper, we use the objective function value of the linear relaxation problem as
the lower bound of the original problem. The linear relaxation problem is derived from
the original problem by converting all binary variables to continuous variables which
changes the constraint si1jt , si2jt , u ivt {0, 1} to 0 si1jt , si2jt , u ivt 1. Table 4
also shows that the difference between the hybrid GA and the lower bound does not
exceed 5% in all problem instances. This provides persuasive evidence that the hybrid
GA performs effectively.
When the size of the problem increases, especially the number of ports, the computation time of the hybrid GA becomes too long. This can be explained as follows.
There are two phases in the hybrid GA. The first is to find the good set of chromosomes

123

784

I.-K. Moon et al.

Table 3 Problem size for 20 randomly generated problems

Problem
instance

Problem size
Number
of ports

Number of variables

Number of
container types

Number
of periods

Integer
variables

Continuous
variables

Number of
constraints
Total
variables

13

1,560

69,264

70,824

16,380

10

13

3,250

170,950

174,200

28,990

13

1,040

32,370

33,410

8,710

13

832

24,544

25,376

7,384

13

2,925

179,361

182,286

32,175
44,096

26

4,576

223,392

227,968

26

5,382

256,230

261,612

47,034

26

1,456

36,816

38,272

11,232

26

2,652

86,580

89,232

21,060

10

26

2,496

69,264

71,760

17,160

11

39

3,510

129,480

132,990

34,320

12

39

1,287

22,815

24,102

8,073

13

39

3,315

113,295

116,610

30,225

14

39

4,212

155,844

160,056

37,440

15

39

5,733

273,273

279,006

58,695

16

52

2,028

42,588

44,616

14,820

17

52

2,704

61,360

64,064

18,928

18

52

8,320

297,856

306,176

60,736

19

52

3,120

85,904

89,024

26,000

20

52

5,616

207,792

213,408

49,920

and it is performed by GA operations, while the second is to solve the LP part of each
chromosome of this good set by using optimization software. The computation time
in the first phase is short but that in the second phase is too long. Since the time to
solve an LP problem has a polynomial relationship with the size of the problem, a
large amount of time is required when the size of the problem is large. For example, in
problem instance 6, the time to solve the LP part (i.e. the second phase of the hybrid
GA) accounts for more than 70% of the total computation time, whereas it accounts
for approximately 81% in problem instance 2 and 88% in problem instance 7. Moreover, the number of continuous variables is directly proportional to the cube of the
number of ports, and the number of constraints is directly proportional to the square
of the number of ports. Therefore, the number of ports has the greatest effect on the
computation time. Throughout the experiments on the above 20 problem instances, it
can be seen that the solution found in the second stage reduces the objective function
value about 4%. However, as we mentioned above, the second stage takes too much
computation time. Hence, if the problem size is larger than those problem instances,
only the first stage should be used and we may lose about 5% in comparison with
optimal solution.

123

Positioning empty containers among multiple ports

785

Table 4 Comparison between MIP, the LP-based GA, and the hybrid GA for 20 problems
Problem instance Computation time (s)

Penalty (%) of the Hybrid GA

MIP

LP-based GA Hybrid GA vs. Lower Bound vs. MIP vs. LP-based GA

1,221

3.10

14,159

4.11

442

34,076

589

2.07

1.96

0.42

182

19,166

358

2.07

1.94

0.22

7,559

4.29

12,072

2.66

27,646 37,001

34,603

2.78

745

1.66

1.54

0.13

1,987

2.61

10

1,396

2.03

11

2,643

2.53

12

18,333 17,132

377

1.57

1.45

0.15

13

2,231

2.25

14

3,837

2.44

15

10,129

2.59

16

38,735

686

2.07

0.20

17

1,084

2.01

18

19

1,513

2.06

20

5,058

2.12

5 Conclusions
In this paper, we considered the problem of empty container positioning with respect
to leasing and purchasing. The problem considers many decision variables, such as
the number of empty containers to be positioned, the number of leased and purchased
containers, and the number of leased containers to be returned, and the solution is
useful to a shipping company for decision making at an operational level. MIP was
proposed for this problem and two GAs were suggested. The results showed that the
LP-based GA and the hybrid GA are capable of solving problems of larger size. Moreover, through experiments, it was demonstrated that the hybrid GA is more efficient
than the LP-based GA in terms of computation time. The numerical experiments also
pointed out the limitation of the hybrid GA when the number of ports becomes too
large.
To solve a very large problem, we should reduce the size of the LP problem. Assumptions may be given to reduce the number of continuous variables and constraints. The
returning of leased containers is an example. If it is assumed that the leased containers can be returned at any port, then the index k can be omitted. Note that k appears
f lk
k
k
in many variables. These are xielk
jvt , x i jvt , wivt , and LIivt . Therefore, omitting k can

123

786

I.-K. Moon et al.

considerably reduce the number of variables. It would also remove constraints relating
to this index [Constraints (2a) and (6)]. Another assumption is not to consider the vessel management. With this assumption, variables and constraints that relate to vessel
transportation are omitted, such as f i1jvt , f i2jvt , si1jt , si2jt , and Constraints (7) and (8).
However, these assumptions make the problem less meaningful and impractical. A
better approach would be to study other heuristic methods that can lead to a feasible
solution for the problem, and this will be the subject of further research.
Acknowledgments The authors are grateful to the careful and constructive reviews from the guest editor
(Professor Herbert Kopfer) and two anonymous referees. The authors also thank Mr. Fischer at Softship
Data Processing Pte. Ltd. for providing practical advices on parameters for the mathematical model. This
work was supported by the Grant of the Korean Ministry of Education, Science and Technology (The
Regional Core Research Program/Institute of Logistics Information Technology).

References
Crainic TG, Dejax P, Delorme L (1989) Models for multimode multicommodity location problems with
interdepot balancing requirements. Ann Oper Res 18:279302
Crainic TG, Delorme L (1993) Dual-ascent procedures for multicommodity locationallocation problems
with balancing requirements. Transport Sci 27:90101
Crainic TG, Gendreau M, Soriano P, Toulouse M (1993) A Tabu search procedure for multicommodity
location/allocation with balancing requirements. Ann Oper Res 41:359383
Gen M, Cheng R (2000) Genetic algorithms and engineering optimization. Wiley, New York
Gendron B, Crainic TG (1997) A parallel branch-and-bound algorithm for multicommodity location with
balancing requirements. Comput Oper Res 24:829847
Gendron B, Potvin JY, Soriano P (2003) A Tabu search with slope scaling for the multicommodity capacitated location problem with balancing requirements. Ann Oper Res 122:193217
Goldberg D (1989) Genetic algorithms in search, optimization and machine learning. Addison-Wesley,
USA
Li JA, Liu K, Leung SCH, Lai KK (2004) Empty container management in a port with long-run average
criterion. Math Comput Model 40:85100
Li JA, Leung SCH, Wu Y, Liu K (2007) Allocation of empty containers between multi-ports. Eur J Oper
Res 182:400412
Lou YC, Guinard M, Chen CH (2001) A hybrid approach for integer programming combining genetic
algorithms, linear programming and ordinal optimization. J Intel Manufact 12:509919
Nemhauser GL, Wolsey LA (1988) Integer and combinatorial optimization. Wiley, USA
Pedroso JP (1998) An evolutionary solver for linear integer programming. BSIS Technical Report No. 987
Rajasekharan M, Peters BA, Yang T (1998) A genetic algorithm for facility design in flexible manufacturing
system. Int J Prod Res 36:95110
Shen WS, Khoong CM (1995) A DSS for empty container distribution planning. Decis Support Syst 15:75
82

123

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.